Click to follow
Text｜Hao Xin
Editor｜Liu Yuqi
The past WAIC (World Artificial Intelligence Conference) has become a large-scale model manufacturer's achievement reporting conference.
The Baidu Wenxin large-scale model was upgraded to version 3.5, which increased the training speed by 2 times and the inference speed by 30 times; Huawei Cloud released the Pangu large-scale model 3.0, including the L0 basic large-scale model, the L1 industry large-scale model and the L2 scene model;
Large-scale models have grown from scratch. At the current stage, even manufacturers standing on the same starting line have embarked on different forks: some are making general-purpose large-scale models, while others are making industrial large-scale models; some are escorting the safe operation of large-scale models, and some are providing tools for making large-scale models.
On July 7, in the "2023 Large Model and AIGC Industry Map" released by the China Academy of Information and Communications Technology, the large model and the upstream and downstream of the AIGC industry chain are divided into four main parts: industry application, product service, model and tool, and basic layer.
(Source: China Academy of Communications)
From the picture, a large model is like building a house, and it is difficult for one company to complete all the links. In other words, just like the era of great exploration of the Internet, everyone has the opportunity to build a large-scale model ecology.
Just as Wu Yunsheng, vice president of Tencent Cloud, head of Tencent Cloud Intelligence, and head of Youtu Lab, told Guangcone Intelligence: "In the era of large-scale models, openness is a very important feature. Large-scale models need to be combined with industry implementation, which requires a lot of cost. In this case, in order to maximize value, only openness. By allowing experts from all walks of life and people with various roles to join in, the entire ecosystem can be made healthier and more possibilities will be created."
While developing, problems are gradually exposed. Compared with the mature large-scale model market in foreign countries, China has not yet built a complete large-scale model industry chain. There are deficiencies in underlying data, chips, and computing capabilities, and it is still very weak in model training and deployment.
Aiming at the pain points in the large-scale model industry chain, Tencent Cloud MaaS large-scale model select store upgrades the technology base, releases the vector database and Xingmai network, and innovates the application scenarios of large-scale models in the industry.
Tracing back to the source, it can be seen that Tencent continues the thinking of the Internet era. Tencent Cloud still does not make general-purpose large models, but continues to make toolboxes and connectors.
Always Toolboxes and Connectors
According to Light Cone Intelligence, as early as June 19, Tencent Cloud announced the industry's large-scale model technology solution. The solution relies on the Tencent Cloud TI platform to build a selected store of large-scale models in the industry, providing customers with one-stop MaaS services. Customers only need to add their own unique scene data to quickly generate exclusive models, and develop low-cost, high-availability smart applications and services based on actual business scene needs.
The Tencent Cloud MaaS large-scale model store highlights two features, one is specialized, and the other is flexible.
"Fine specialization" is mainly reflected in the training of industry models. Tencent Cloud has built large models of multiple industries such as finance, government affairs, cultural tourism, media, and education in its technology base.
To use an analogy, it is like that college students are assigned to different majors as soon as they enter school, and then continue their postgraduate and doctoral studies on this basis. The same is true of Tencent Cloud's thinking. The training data is first honed in the industry's large model, and then combined with the company's private data, fine-tuned to generate a company-specific model. It can be said that industry cognition runs through the entire process of model training, reasoning, and deployment, thereby improving the application capabilities of industry scenarios.
"Flexibility" is mainly reflected in the enterprise's ability to call and use models and tools. In the Internet era, Ma Huateng once positioned Tencent's role as a "toolbox". In the era of large-scale models, the role of the Tencent Cloud TI platform is similar.
The tool chain determines whether an enterprise can combine large-scale model capabilities with its own business and products. To this end, Tencent Cloud provides large-scale model toolboxes and supporting services including data labeling, training, evaluation, testing, and deployment. Enterprises can select and combine tools locally as needed, conduct privatized data training under the premise of ensuring security, and can also customize model services with different parameters and specifications according to business scenario requirements.
In the past, Tencent connected B-end merchants and C-end users, and now Tencent Cloud has also replicated this capability to large-scale model stores. The tool chain formed by the combination of tools is only a link in the platform, connecting enterprise products and large models; the other main line - data (private + public), connects large models, enterprises, industries and users.
Industry application is the starting point and the end point. As Wu Yunsheng said, "No matter what kind of technology, our most fundamental starting point is to solve practical problems."
Large Model Accelerator
Tencent Cloud seems to have figured out how to find a rhythm that suits you on the large-scale model track.
Tang Daosheng, Senior Executive Vice President of Tencent Group and CEO of the Cloud and Smart Industry Business Group, once said: "The key is to do a good job in the underlying algorithms, computing power and data, and the most important thing is to implement the scene."
Continuing with this idea, Tencent Cloud grasped the underlying algorithms, computing power and data with one hand, and implemented the scene with the other hand. The MaaS large-scale model store has achieved a comprehensive upgrade.
In terms of technical foundation, Tencent Cloud focused on the word "fast", released the Xingmai network and vector database, and installed "Hot Wheels" for the large model.
Large-scale models have entered the era of trillions of parameters, and the computing power of a single server is limited. It is necessary to connect a large number of servers through a high-performance network to build a large-scale computing power cluster.
Based on this, Tencent Cloud overcomes the problem of computing power loss in large cluster scenarios through comprehensive optimization of processors, network architecture, and storage performance, and officially released a new generation of HCC (High-Performance Computing Cluster) high-performance computing clusters.
The cluster uses Tencent Cloud Xinghai's self-developed server, equipped with Nvidia's latest generation H800 GPU, which can increase GPU utilization by 40%, save 30%-60% of model training costs, and bring a 10-fold improvement in communication performance for large AI models. Based on Tencent Cloud's new-generation computing power cluster HCC, it can support a super-large computing scale of 100,000 cards.
According to Tencent Cloud, the computing performance of Tencent Cloud's new generation cluster is up to three times higher than that of the previous generation, making it the most powerful large-scale computing cluster in China.
The high-performance computing group is a basic capability, and the application of its technology reflects the cost reduction and efficiency increase through technical means.
First of all, compared with a large number of scattered computers, high-performance computing clusters can reduce hardware costs and operation and maintenance costs, and at the same time facilitate centralized management. Second, it can improve the efficiency of calculation and search. Provides distributed computing capabilities to support vector databases; it can also perform complex scientific calculations and modeling, which is why Tencent Cloud's "AI for Science" can quickly achieve results in astronomy and oracle bone inscriptions.
During the training process of large models, Tang Daosheng talked about the problem of data quality. He said: "At present, general large models are generally trained based on extensive public literature and network information. The information on the Internet may contain errors, rumors, and biases. Many professional knowledge and industry data are insufficiently accumulated, resulting in insufficient industry pertinence and accuracy of the model, and excessive data noise."
The significance of data to large model training is self-evident. At present, in addition to excessive data noise, there are also many problems in data processing, data update, and data security.
In addition, there is a fatal shortcoming of the large model - there is no long-term memory, and the C-side dialogue scene can also be asked again, but it may cause the system to crash when it is applied in the industry.
OpenAI realized this problem very early. By cooperating with Zilliz, Pinecone, Weaviate and other vector database companies, it configured ChatGPT with an "external cache". The vector database + large model is also called the "golden partner".
The popularity of foreign vector databases has spurred the acceleration of domestic manufacturers, and Tencent Cloud has also caught up with the first wave, releasing the first domestic AI native vector database.
For large model scenarios, it realizes comprehensive AI in the access layer, computing layer, and storage layer:
At the access layer, intelligence supports direct retrieval of natural language texts;
At the computing layer, AI operators are used instead of enterprises to find/tune AI algorithms, shortening the access period from one month to 3 days;
At the storage layer, intelligent compression algorithms are integrated to reduce vector storage costs by 50%.
Enterprise data access needs to be divided into three steps, namely text segmentation, vectorization, and import. In the past, these three steps were performed by different companies, so the cycle was infinitely long. However, Tencent Cloud turned the three steps into one step, directly realizing one-stop access, and improving the efficiency by 10 times.
However, from the perspective of parameters, the performance of Tencent Cloud's vector database is still in its infancy.
For example, Tencent Cloud Vector Database supports up to 1 billion-level vector retrieval scale, and controls the delay at the millisecond level. As a comparison, Milvus can support a maximum vector retrieval scale of 56 billion, and supports millions of vector similarity searches per second.
But 1 billion levels can also be said to be an entry-level parameter for vector databases. Pinecone's official demo shows that it can search in 1 billion vectors in real time; the Weaviate algorithm can support vector indexes on the order of one billion.
If a worker wants to do a good job, he must first sharpen his tools. Step by step from the bottom to solid technology, it seems that Tencent Cloud has taken a slow road, but Xiaobu is going fast, and after realizing rapid iteration, it will drive the improvement of the entire ecosystem.
50 scenes, Tencent Cloud landing in batches
Scenarios have always been the product culture that Tencent emphasizes, that is, to make a product or launch a function, the first consideration is whether you can find the scene and find the users.
Also cutting into MaaS, standing on the new starting line, Tencent Cloud aims at the rigid needs of enterprise applications through the accumulated industry Konw-how, promotes the implementation of large-scale model applications, and uses the scene as a sharpening stone for training large-scale models.
"Although the large model is good, it still has a high threshold for use. Especially for some enterprises in traditional fields, the general-purpose large model cannot be accurately adapted to meet the expectations of reducing costs and increasing efficiency. What enterprises need is to truly solve a certain problem in actual scenarios, rather than solve 70%-80% of the problems in 100 scenarios." Wu Yun said.
Tencent Cloud believes that large-scale models are not just a game for a few people. Transforming large-scale models from "generalists" to "specialists" may be a feasible path for enterprises. The role of Tencent Cloud is to lower the threshold and provide one-stop services to help enterprises skip the "cold start" stage of model training and deployment.
According to Lightcone Intelligence, based on Tencent's long-term accumulation in the Internet industry, Tencent Cloud has joined hands with more than ten industry leaders such as finance, cultural tourism, government affairs, media, and education to jointly create more than 50 large-scale industry model solutions. These are the key service industries of Tencent CSIG.
In the financial risk control scenario, Tencent Cloud’s risk control model integrates Tencent’s past 20 years of black and gray industry confrontation experience and thousands of real business scenarios to provide a financial risk control solution. Based on the prompt mode, enterprises can iterate risk control capabilities, from sample collection, model training to deployment and launch, and realize zero manual participation in the whole process. At present, the modeling time has been reduced from 2 weeks to only 2 days.
In the interactive translation scenario, based on the industry's large model technology, Tencent Cloud does not require millions of training data, and can obtain good translation results by using small sample training, so that each interactive translation can play a real-time role in improving the translation of the next sentence.
Taking industry scenarios as the starting point, technology and applications are iteratively upgraded at the same time, which is obviously faster and more effective.
According to the latest data from Tencent Cloud in WAIC, the above-mentioned financial risk control solution has improved the efficiency by 10 times compared with the previous one, and the overall anti-fraud effect has increased by about 20% compared with the traditional model; Tencent Cloud Digital Homo sapiens factory has more than 10 built-in AI algorithm models.
Technology and application run on two legs at the same time, the underlying large model supports the landing of the application scene, and the scene in turn feeds back the large model.
Just as Tencent Cloud emphasizes that "industrial scenarios are the best training ground for large models", the industry experience learned in the early stage of large models can be corrected in real-world application scenarios, and once again deposited on the base of the industry model on the Tencent Cloud MaaS platform, renewing cognition, and repeating this cycle, industry large models will become more refined and enterprises will use them more effectively.
On the other hand, mature application landing scenarios may open up a new path for the commercialization of large models.
Facts have proved that coveting the excitement for a while is not long-term. Even ChatGPT, which has absolute technical barriers, is facing the fate of declining traffic. The founder of OpenAI even bluntly said that the reason why the ChatGPT plug-in is not as expected is that people want to use the capabilities of GPT in their own applications.
Opening the next stage of competition for large models, scenarios and commercialization capabilities may become closer.
Just as Wu Yongjian, vice president of Tencent Cloud and head of Tencent Cloud Intelligent R&D, believes: "The Internet has entered the stage of being purely free at the beginning, and gradually moving to the stage of how to commercialize certain scenarios. This is not brought about by the big model, but the way for us to be commercialized by the big model has become clearer."
Welcome to pay attention to the "Light Cone Intelligence" CSDN number, pay attention to cutting-edge technology!